Kepler’s cover photo
Kepler

Kepler

Software Development

New York, NY 1,852 followers

The deterministic substrate for AI.

About us

We're building the ground truth platform for AI. AI is great at understanding what you're asking for. It's terrible at giving you answers you can trust. Everyone else is trying to make AI more accurate. We made accuracy the only possible outcome by building a platform that separates what AI does well from what code does well. AI handles the conversation. Code handles the truth. The result is the first AI system that can show its work. We automatically ingest data from scattered sources, structure it into a unified platform, and deploy specialized AI agents that conduct deep research with full transparency. Every insight links back to authoritative sources. Every conclusion reveals its reasoning. Every answer can be defended with complete confidence. We're starting in finance, where being wrong costs millions and speed wins deals, but we're building the foundational data layer for the AI era, applicable anywhere that decisions depend on trustworthy data. We're a team of ex-Palantir engineers who built data infrastructure for the world's most demanding organizations. We've raised from top investors, we're working with partners who need this yesterday, and we're building a small team obsessed with foundational technology. If you want to build the platform that becomes the ground truth for AI, we're just getting started.

Website
https://kepler.ai
Industry
Software Development
Company size
2-10 employees
Headquarters
New York, NY
Type
Privately Held
Founded
2025
Specialties
AI, Data Platform, and Data Security

Locations

Employees at Kepler

Updates

  • Kepler reposted this

    PwC sells advice on adopting AI responsibly. It just published four reports where the AI made things up. One promotes a product called Citizen Pulse. Citizen Pulse does not exist. It names Denmark, Saudi Arabia, the United States and Australia as customers. The footnotes proving those governments use it link to those governments' homepages. GPTZero puts the report at 84% likely to be entirely AI generated. PwC says it "is updating a limited number of supporting citations." Deloitte refunded a government. EY withdrew a study. KPMG pulled a report. All four of the Big Four have now published hallucinations. All four also sell AI governance advice. Matt Levine's read is that these reports were never meant to be read. Nobody cares enough to write them, so AI writes them, so they are full of nonsense, and it does not matter, because the only job of the report is to signal that PwC exists and has thoughts about AI. Which is exactly what makes this expensive. The document had one job. It was a credential. It is now the best available evidence that the firm selling AI governance cannot run its own. You do not buy that back with a citation update. Kepler does not do audit. Yet. But if PwC wants a research tool that can prove its own sources exist, we'll stand up an audit vertical by Friday. The first design partner slot is yours.

    • No alternative text description for this image
  • Kepler reposted this

    AI Engineer just posted the video from my talk at the AIE World's Fair. Every AI company is hiring forward deployed engineers right now. Almost all of them are hiring them into sales. That is not where this started. In my talk, I share stories about my actual experience being an FDE. Like the time we took on a shipping customer who opened with a 47-page requirements doc and a three-month estimate. We spent four months scoping it. Then someone finally showed up on site and asked the dispatcher what they do first on Monday morning. They check whether trucks are late. We shipped a Slack alert in four hours. Whoever defines the problem owns the solution. This is how we think of FDE at Kepler. Not as a sales team. As the product team. If you are early stage, FDE is not your go to market. It is your product strategy.

    View organization page for AI Engineer

    27,192 followers

    AI can write the code. It still cannot own the outcome. That gap is where forward deployed engineers work: between a model that can build almost anything and a customer who needs one specific thing to work inside a messy real company. Each company gives the role a different shape, but the mandate is the same: turn AI capability into a result the customer can measure. Our full Forward Deployed Engineering Track from AI Engineer World's Fair is live now on YouTube. Full track: https://lnkd.in/eGzEjzUG Vinoo Ganesh, CEO and Co-Founder of Kepler, spent seven years doing this at Palantir and later ran the rotation that turned software engineers into forward deployed engineers. His argument is that FDE was never a sales motion. It is product strategy in disguise. Watch what customers actually do, ship the smallest thing that changes the outcome, and treat every repeated task as a missing product feature. Kevin Bai of Anthropic, formerly at Palantir and the founding FDE at Rippling, starts from the business model. When the buyer is not technical and the product cannot be configured off the shelf, you are not selling software or services. You are selling the outcome between them. That can land contracts a self-serve motion never reaches, but only if the custom work compounds into a reusable platform instead of becoming nine different products for nine customers. Nat Meurer, Head of Agent Engineering at Sierra, tells the dirty secret: forward deployed engineering has never meant one job. At Palantir it stretched across DevOps, data integration, ontology work, and solution architecture. As coding agents make software cheaper to write, the durable part is everything around the code: understanding the customer, integrating the data, and being accountable to an outcome. Her conclusion is that agent engineering is FDE reborn. Jia Wu, Deployed Engineering Lead at Cognition, shows how to measure that outcome. Most of the coding agent market optimizes for token usage. Her team measures before Devin lands and again after it is fully activated, reporting an 82% reduction in the targeted work. One team adopting an agent is a step function; a whole enterprise adopting it can be parabolic. Also on the track: - Eno Reyes, Factory - Pauline Brunet, Cursor - Vasuman Moza, Varick Agents - Sunny Rekhi, Decagon - Leo Mehr, Ramp From six week Cursor deployments and air gapped Factory agents to Decagon upstreaming custom integrations and Ramp turning customer intake into agent written specs: this is a full day on the engineering discipline between AI capability and customer outcome. Full track: https://lnkd.in/eGzEjzUG

    • No alternative text description for this image
  • Our CEO, Vinoo Ganesh spoke at the AI Engineer Conference about how verifying AI work has not gotten any easier and where the solution lies.

    Last week, I was invited to speak at the AI Engineer conference about the one thing three years of model progress hasn't made easier. AI made producing work limitless. It did nothing to make verifying that work any easier. You cannot look at an answer and tell whether it is true. That information lives somewhere else, in the record of how the answer was made. Software has that record. Every commit, every pull request, every review, kept in perpetuity. The work product is the proof. Finance has nothing like it. An analyst types a number out of a 10-K into a spreadsheet and the reasoning evaporates. We solve this by building the record. The model is never allowed to write a number, because a number it generates has no history, and a number without history cannot be checked. Correct and uncheckable are the same thing to whoever signs the memo. The model points instead. Filing, page, line item. Code retrieves the value. A deterministic check confirms it against the source, one input and one right answer, with no second model offering an opinion on the first. Anything that fails the check never reaches the analyst. Thanks to the AI Engineer team for the stage. If these are the kind of problems you want to be working on, get in touch, we're hiring.

    • No alternative text description for this image
  • View organization page for Kepler

    1,852 followers

    Founding Engineer Christopher Martin writes about some of the surprising challenges associated with extracting data from tables

    Three frontier models, one clean table, one column average. Three different answers, and not one of them was bad at arithmetic. They just couldn't agree on how to handle the blank cells. A table's layout is built for reading, not computing, and most financial AI plays along: fluent, confident, inconsistently wrong. I just published a post on the harder path. You compute on a copy of the table, but you keep the original and trace every number back to the exact cell it came from, because you can't be your own evidence. Proof has to come from the one thing in the analysis you didn't create. That's where most LLM wrappers fall apart. Kepler is verifiable, not just plausible. Link in comments.

  • View organization page for Kepler

    1,852 followers

    What happens when there are hundreds different models for different use cases? The model stops being scarce and there are new limiting factors.

    This week, two announcements quietly said the same thing: the model has stopped being scarce. Microsoft shipped its own family of in-house models, openly framed as reducing its dependence on OpenAI. DeepSeek is reportedly raising north of $50B, the same lab that trained a frontier model on a fraction of the usual budget. The giant that used to rent its intelligence is now building it, and an upstart proved you don't need a giant's budget to reach the frontier. Read this as one story: models are becoming abundant. Some general and powerful, some small and tuned for one job, and the supply keeps growing. You no longer pick a winner, you pick the best one for the task in front of you. And the workflows built on top are going the same way, because the models are absorbing them faster than anyone can build them. Here is the part nobody has claimed. A smarter model is still a confident guess, and more of them doesn't make a single answer more trustworthy. In finance, law, medicine, the constraint was never intelligence. It was whether you could prove the answer. That proof lives in a layer underneath the model: the data pipelines, the memory, the provenance that makes an output defensible. That is the infrastructure we are building at Kepler.

  • View organization page for Kepler

    1,852 followers

    Welcome to the team Michelle!

    Happy to share that I've joined Kepler. Excited to be working with a talented team, founded by Vinoo Ganesh and John McRaven on the challenge of making AI systems more reliable and traceable, starting first with finance. We're building the deterministic infrastructure for AI so outputs can be verified, and trusted in real-world workflows. If you're working on AI, agentic systems, or just enjoy solving messy data/finance problems, let's chat!

    • No alternative text description for this image
  • Kepler reposted this

    Everyone's talking about being AI-pilled. At Kepler, we're focused on making a different pill: the pain killer for the headaches financial analysts actually live with every day, and night. A few weeks ago, I asked a simple question on LinkedIn: if AI is so advanced, why are analysts still at their desks at 4AM? I received an overwhelming number of messages from people in finance saying the same thing. The tools they have today are supplements. Nice to have, maybe good for you, but they don’t help with the pain. It is extremely difficult to verify results LLMs give you. As the task gets more complicated, the combination of lack of verifiability and increased hallucination makes LLMs nearly impossible to be truly adopted in an analyst’s work. That's what we've built Kepler to fix. A verifiable and deterministic AI infrastructure where LLM and code never cross and you get results that are not only accurate, but can be verified with one click. Not probably right. Provably right. We have started in finance and are building towards solving the pain in many more industries. I just published a blog on Kepler.ai with the recent GameStop - eBay deal as an example on how Kepler is used to speed up workflows that many financial analysts, myself included, used to dream about getting help on all the time. Link in comments.

    • No alternative text description for this image
  • Kepler reposted this

    Today is a historic day for the US stock market. SpaceX priced the largest IPO ever: a $75 billion raise at $135 a share, valuing the company at $1.77 trillion. Over $250 billion in total investor orders, 4x oversubscribed. Retail investors alone placed more than $100 billion in orders, and Elon Musk could become the world's first trillionaire before the year is out. Ahead of the open, we used Kepler to build a full model from the S-1: 3-statement build, segment breakdowns (Space, Connectivity, AI), capital structure, KPIs. It's set up to roll forward when the first public quarter prints. By hand, a model this size takes an experienced analyst three full days. Most of that is reading, hunting for numbers, and retyping them. With Kepler it took us a couple of hours, and most of that was thinking, not typing. Kepler pulled the data straight from the S-1, and with our deterministic infrastructure, every number and claim in the model is cited to the exact page and sentence it came from. No black box, no blind trust in AI. We are sharing the model for free. Link in the first comment.

    • No alternative text description for this image
  • Kepler reposted this

    Today is a historic day for the US stock market. SpaceX priced the largest IPO ever: a $75 billion raise at $135 a share, valuing the company at $1.77 trillion. Over $250 billion in total investor orders, 4x oversubscribed. Retail investors alone placed more than $100 billion in orders, and Elon Musk could become the world's first trillionaire before the year is out. Ahead of the open, we used Kepler to build a full model from the S-1: 3-statement build, segment breakdowns (Space, Connectivity, AI), capital structure, KPIs. It's set up to roll forward when the first public quarter prints. By hand, a model this size takes an experienced analyst three full days. Most of that is reading, hunting for numbers, and retyping them. With Kepler it took us a couple of hours, and most of that was thinking, not typing. Kepler pulled the data straight from the S-1, and with our deterministic infrastructure, every number and claim in the model is cited to the exact page and sentence it came from. No black box, no blind trust in AI. We are sharing the model for free. Link in the first comment.

    • No alternative text description for this image

Similar pages